A Novel GAN-Based Anomaly Detection and Localization Method for Aerial Video Surveillance at Low Altitude
, Irene Cannistraci 1, Marco Cascio 1, Luigi Cinque 1, Anxhelo Diko 1, Alessio Fagioli 1, Gian Luca Foresti 2, Romeo Lanzino 1, Maurizio Mancini 1, Alessio Mecca 2 and Daniele Pannone 1
Round 1
Reviewer 1 Report
Major points:
1. The abstract of the paper should briefly explain the summary of this work. In the said paper authors should add some motivation and add the dataset information.
2. What are the authors' scientific contributions? The utilization of GAN models is not innovative enough. Many studies have proposed such models and it’s existing technology. Would you mind clarifying them?
3. “Discussion” section should be separated and explained in a more highlighting, argumentative way. The author should analysis the reason why the tested results are achieved.
4. How were the hyperparameter values of the P-GAN architecture chosen? Were they through a grid search? If they were through a grid search, the authors could improve the table by providing the ranges of different hyperparameters they have considered for a grid search.
5. The current challenges are not crystal clearly mentioned in the introduction section of this paper. I suggest adding a dedicated paragraph about the current challenges of this area followed by the authors’ contribution to overcoming those challenges.
6. The contributions in the current version are not innovative enough and very weak. I strongly recommend improving the contributions in the revised version by adding some detailed info.
7. The manuscript, however, does not link well with recent literature on recognition that appeared in relevant top-tier journals, e.g., the IEEE Intelligent Systems department on “Advanced Fusion-Based Speech Emotion Recognition System Using a Dual-Attention Mechanism with Conv-Caps and Bi-GRU Features” are missing it should be cited.
8. The authors must provide clearer details regarding the research findings, including suitable figures and tables, in its present form, it could hardly be understood what the authors explored. Figures 2, 3, and Tables 2-4 really say minor. Meanwhile, the authors have also
9. The comparison table is missing. The authors only compare the results among recent SOT. The authors must compare other existing and popular deep learning models such as various sequential models.
1. The linguistic quality needs improvement. It is essential to make sure that the manuscript reads smoothly- this definitely helps the reader fully appreciate your research findings.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
the article is interesting, I think it will be necessary to focus on the conclusion the importance of the work developed and make a comparative table between this study and other similar studies
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
In this paper, the authors propose a novel unsupervised GAN-based anomaly detector. The two branches structure is interesting and seem works well in the anomaly detection. However, some details should be explained to facilitate the reader's understanding, here are some comments about this paper:
1. What is the relationship between the detector branch and the localizer branch? It seems that the detector branch is just trained to expand samples.
2. What is the training and testing flow of this network? Are the recognition and localization branches trained separately or are they the same part, for instance, the generator and discriminator are trained simultaneously?
3. If the reviewer correctly understands this model, locator branch only works if the detector notices some anomalies. If so, the detector branch should have a threshold for this work, how do authors find a define it?
4. How did the GAN structure (localizer branch) localize the anomalous elements? Compared with the single branch methods, what is the advantage of the two branches structure?
5. In equation 4 and equation 7, how did their weights define? Are they defined as constants by experience or learned by the model?
6. How efficient is this approach? For example, its cost of inference time?
7. Can you show me a graph of the changes in the loss of the generators and discriminators in this model?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors successfully addressed my comments and suggestions. Good Luck!
Reviewer 3 Report
The author has addressed all my concerns and I have no further comments.
